Google DeepMind unveiled Genie 4, and the update adds native audio and human speech. It also pushes the world model closer to real-time simulation, with more realistic physics and a planned rollout soon for developers and Google AI Ultra subscribers.
The system generates experiences frame by frame as a user explores them. It also uses Gemini 3.8 Flash as a guiding agent to break down goals and evaluate each output frame, which gives DeepMind a tighter control loop but still leaves the model short of long-run persistence.
Genie 4 and Gemini 3.8 Flash
DeepMind says Genie 4 improves on Genie 3 in almost every way, and that matters because the company has described world models as a key element in the march toward AGI. The practical change for users is not abstract ambition; it is a system that can adjust weather, add elements, or move the camera while the simulation keeps running.
Genie 4 also reaches a new level of realism in physics. DeepMind says it handles aerodynamics and collisions better, and that people in the simulation should look less uncanny while moving and reacting more naturally to the changing environment.
Native audio in Genie 4
Genie 4 is the first DeepMind world model with native audio. Its videos will include ambient background sounds, and users can specify audio elements in a prompt, which makes the simulation feel less like a silent demo and more like a place that can respond to direction in more than one channel.
DeepMind says Genie 4 will also be able to create interactive dialog that adapts as the simulation changes. That is the feature that matters for developers trying to prototype interactive scenes, because the audio layer is no longer just decoration around the visuals.
Developers and Google AI Ultra
DeepMind has been working with a small group of trusted testers to build toward a wider release. Google plans to roll out Genie 4 for developers via the Gemini API and AI Studio soon, while Google AI Ultra subscribers sit on the same early-access path.
The catch is persistence. DeepMind says Genie 4 worlds still remain consistent for only a couple of minutes before they forget how things started, so the model can simulate a moment but not yet sustain a long story without losing state. That is the limit readers should watch, along with pricing, which Google did not give.
